Models Round Table

Answer one task with several models instead of one. Each model answers the raw prompt blind — you, the orchestrator, add no analysis, framing, or preferred answer before they respond, so the seats stay independent and unbiased. Then a dedicated organizer maps the answers into a five-dimension structured analysis, one bounded gap-repair round closes the open points, two judges validate the still-open ones, and a dedicated synthesizer writes the single consensus answer. The deliverable is that consensus answer plus a report of how it was reached.

The pipeline is shaped by where multi-model gains actually come from: the synthesis step is the dominant lever (it is given to a real model, not hand-stitched by the orchestrator), and model diversity is the secondary lever (kept as the default, but not mandatory).

Read-only: seats, organizer, judges, and synthesizer answer and analyze; they do not modify the repo. (Turning a consensus into code is the job of feature-models-roundtable / implement-and-review.)

Hard Rules

1. You are the moderator, not a seat. Add no analysis, hints, or preferred answer before the seats respond — the opening fan-out must be blind so none of your bias leaks in. Never count your own answer as a vote. You do not write the consensus answer either — a synthesizer model does; you validate it against the record.
2. Read-only — no mutation. Seats, organizer, judges, and synthesizer produce answers/opinions only; they never edit files, run mutating commands, or implement anything. Read-only information tools (web search/fetch, repo reads) are allowed only under an explicit shared tool profile (see Preflight) — never write/exec tools, never by default.
3. Blind opening. Every seat gets the raw prompt (plus read-only repo context only when the task is about this repo), no peer answers, no "right" answer.
4. Bounded reconciliation. At most one gap-repair round, then a two-judge panel, then your final call. Do not loop indefinitely.
5. Identical conditions across seats. Every active seat in a run gets the same tool profile, the same read-only sources, and the same budget. Uneven tool access biases the panel.
6. Never fabricate a seat, judge, or diversity. Missing CLI/tool → drop the seat, lower confidence, and say so. Pass --disable-fallback to every runner so none silently borrows another provider. Duplicate (self-paired) seats must be labeled as same-model samples and must lower diversity confidence — never report them as genuine model diversity.

Seats